Estimating Dynamic Programming Models

Transcription

1 Estimating Dynamic Programming Models Katsumi Shimotsu 1 Ken Yamada 2 1 Department of Economics Hitotsubashi University 2 School of Economics Singapore Management University Katsumi Shimotsu, Ken Yamada DP models JEA 2011 tutorial session 1 / 42

2 Dynamic Programming (Structural) Models For assessing various policy proposals (e.g. pension reform, export subsidy), understanding the dynamic response of individuals and firms is important. Regression models are subject to the Lucas critic. It is desirable to model dynamic optimizing behavior on individuals/firms explicitly, along with how the state of economy evolves. Katsumi Shimotsu, Ken Yamada DP models JEA 2011 tutorial session 2 / 42

3 Dynamic Programming Models: Examples Rust (1987): bus engine replacement Rust and Rothwell (1995): nuclear plant operation Rust and Phelan (1997): retirement decision and pension/health plan Keane and Wolpin (1997): schooling, work and occupational choice Imai and Keane (2004): labor hours choice with human capital accumulation Possible to quantitatively assess the impact of public policies that have never been implemented (counter-factual simulation) Katsumi Shimotsu, Ken Yamada DP models JEA 2011 tutorial session 3 / 42

4 Machine Replacement Model (Rust, 87) Consider a mechanic who maintains a computer. His objective is to maximize the discounted sum of utilities [ ] E β t U(a t, x t, ɛ t ; θ). max a 1,a 2,... t=1 a t {0, 1}: machine replacement decision x t : observable state variable (machine age) ɛ t : state variable observable to the mechanic, but not to econometrician: additional information on machine condition θ: parameter of interest Katsumi Shimotsu, Ken Yamada DP models JEA 2011 tutorial session 4 / 42

5 Machine Replacement Model (Rust, 87) The utility function: U(a t, x t, ɛ t ; θ) = mc(x t ; θ) a t rc(x t ; θ) + ɛ t. mc(x t ; θ): machine maintenance cost rc(x t ; θ): machine replacement cost ɛ t : unobserved (to econometrician) state variable Transition of x t : x t = a t 1 + (1 a t 1 )(x t 1 + 1). The state variable evolves according to (x t, ɛ t ) p(x t, ɛ t x t 1, ɛ t 1, a t 1 ; θ) Katsumi Shimotsu, Ken Yamada DP models JEA 2011 tutorial session 5 / 42

6 Machine Replacement Model (Rust, 87) max a 1,a 2,... [ ] E β t U(a t, x t, ɛ t ; θ), (x t, ɛ t ) p(x t, ɛ t x t 1, ɛ t 1, a t 1 ; θ) t=1 We want to estimate θ from the data {a t, x t },t = 1,..., n. From the mechanic s point of view, he can solve this problem to obtain an optimal decision rule: a = δ(x, ɛ; θ). Because ɛ is unobservable to econometrician, the empirical implication of the model is the conditional choice probabilities (cf. probit model): P(a x; θ) = I {a = δ(x, ɛ; θ)} g(dɛ) Katsumi Shimotsu, Ken Yamada DP models JEA 2011 tutorial session 6 / 42

7 How to compute P(a x; θ) Let s t = (x t, ɛ t ). Define the value function as [ ] W θ (s t ) = max E β s U(a s, s s ; θ) a t,a t+1,... s t. We may compute the value function by finding the fixed point of the Bellman equation { } W θ (s t ) = max a A U(a, s t, θ) + β W θ (s t+1 )p(ds t+1 s t, a). U(a, s t, θ): today s utility W θ (s t+1 ): maximum future utility when tomorrow s state is s t+1 s=t Katsumi Shimotsu, Ken Yamada DP models JEA 2011 tutorial session 7 / 42

8 How to compute P(a x; θ) Bellman equation W θ (s t ) = max a A { U(a, s t, θ) + β } W θ (s t+1 )p(ds t+1 s t, a). When ɛ t is continuously distributed, finding a fixed point of this Bellman equation becomes difficult The space of s t can be very large need to evaluate W θ (s) at many points Numerical integration in computing W θ (s t+1 )p(ds t+1 s t, a) Katsumi Shimotsu, Ken Yamada DP models JEA 2011 tutorial session 8 / 42

9 How to compute P(a x; θ) Conditional Independence Assumption: the transition of x t and ɛ t can be written as p(x t+1, ɛ t+1 s t, a t ) = g(ɛ t+1 x t+1 )f (x t+1 x t, a t ). Any statistical dependence between ɛ t and ɛ t+1 is transmitted entirely through x t+1. The probability density of x t+1 depends only on x t and not ɛ t. ɛ t is a noise superimposed" on x t. Katsumi Shimotsu, Ken Yamada DP models JEA 2011 tutorial session 9 / 42

10 How to compute P(a x; θ) Define the integrated value function as V θ (x t ) = W θ (x t, ɛ t )dg(ɛ t x t ) V θ (x) satisfies another Bellman equation V θ (x) = max a A { U(a, x, ɛ, θ)dg(ɛ x) + β } V θ (x )f (dx x, a). V θ (x) is a fixed point of a separate mapping on the reduced space of x rather than the space of s = (x, ɛ). Under some assumptions on U(a, x, ɛ), one does not need to use numerical integration to compute the right hand side. Katsumi Shimotsu, Ken Yamada DP models JEA 2011 tutorial session 10 / 42

11 How to compute P(a x; θ) Example: U(a, x, ɛ; θ) = u(a, x; θ) + ɛ(a), and ɛ(a) follows extreme value dist n (logit error), independent across a. Then one does not need to use numerical integration to compute the right hand side of the Bellman equation. Further, P(a x; θ) admits a multinomial logit formula P(a x; θ) = exp[u(a, x; θ) + EV θ(a, x)] J j=1 exp[u(j, x; θ) + EV θ(j, x)], where EV θ (a, x) = V θ (x )f (dx x, a) = the maximum future utility when the pair of the current action and observable state is (a, x) Katsumi Shimotsu, Ken Yamada DP models JEA 2011 tutorial session 11 / 42

12 Nested Fixed Point (NFXP) algorithm (Rust, 87) Given a (cross-sectional) data set {a i, x i } n i=1, the NFXP solves Inner loop: for each candidate value of θ, solve the integrated Bellman equation to compute EV θ (a, x) and then compute the equilibrium conditional choice probabilities P(a x; θ). Outer loop: max θ N 1 N i=1 ln P (a i x i ; θ) Then, ˆθ is the MLE. Computational cost depends on the size of the state space (the support points of x and a). Computationally intensive when x may take many different values. Katsumi Shimotsu, Ken Yamada DP models JEA 2011 tutorial session 12 / 42

13 Alternate fixed point mapping Hotz and Miller (1993, ReStud), Aguirregabiria and Mira (2002, ECMA; 2007, ECMA) Computationally attractive alternative to the NFXP These procedures use the fixed point in the space of P(a x): the conditional choice probability of action a given x. Unlike the value function, we can obtain a good guess" of P(a x) from the data, typically from a frequency estimator. Katsumi Shimotsu, Ken Yamada DP models JEA 2011 tutorial session 13 / 42

14 Relation between value function and P(a x) Given the value function, computing P(a x) is not very difficult (logit formula). But computing the value function [ ] W (s t ) = max E β s U(a s, s s ; θ) a t,a t+1,... s t directly from P(a x) is difficult, even with simulations. Assume additive separability: s=t U(a, s) = u(a, x) + ɛ. Katsumi Shimotsu, Ken Yamada DP models JEA 2011 tutorial session 14 / 42

15 Alternate fixed point mapping The Bellman equation W (s t ) = max a A { u(x t, a) + ɛ(a) + β } W (s t+1 )p(ds t+1 s t, a). Define the integrated value function as V (x t ) = W (x t, ɛ t )dp(ɛ t x t ) With additive separability, V (x) satisfies another Bellman equation = V (x) max a A { u(x, a) + ɛ(a) + β } V (x )f (dx x, a) dp(ɛ t x). Katsumi Shimotsu, Ken Yamada DP models JEA 2011 tutorial session 15 / 42

16 Alternate fixed point mapping Mapping from P(a x) to the value function V (x) = { } P(a x) u(x, a) + E[ɛ(a) x, a] + β V (x )f (dx x, a) X a A where E[ɛ(a) x, a] is the expected value of ɛ conditional on action a is chosen. E[ɛ(a) x, a] admits a simple closed-form representation. Combining this with the mapping from the value function to P gives a mapping Ψ(θ, P): P(a x) {a,x} P(a x) {a,x}. True P is the fixed point of this mapping. Katsumi Shimotsu, Ken Yamada DP models JEA 2011 tutorial session 16 / 42

17 Nested pseudo-likelihood (NPL) estimator (Aguirregabiria and Mira, 02, 07) Define P = P(a x) {a,x} : vector of conditional choice probabilities for all a and x. Start from a guess of P, P 0. Typically, a frequency estimator. Estimate θ by θ 1 = arg max Update P by P 1 = Ψ( θ 1, P 0 ). n ln Ψ(θ, P 0 )(a i x i ). Update θ and P further by θ 2 = arg max n i=1 ln Ψ(θ, P 1 )(a i x i ), P 2 = Ψ( θ 2, P 1 ),... i=1 Katsumi Shimotsu, Ken Yamada DP models JEA 2011 tutorial session 17 / 42

18 Nested pseudo-likelihood (NPL) estimator (Aguirregabiria and Mira, 02, 07) Once P is fixed, evaluating Ψ(θ, P) for different θ does not require computing a fixed point. In a single agent model (for example, bus engine model), θ 1 is first-order equivalent to the MLE. θ j approaches to the MLE in a higher-order sense as j increases. (Kasahara and Shimotsu, 2008) In multiple agent (game-theoretic) model, θ 1 is not the MLE. Iterating the updating may give a better estimator. Katsumi Shimotsu, Ken Yamada DP models JEA 2011 tutorial session 18 / 42

19 Game-theoretic models We may view the bus engine model as a model with a fixed point constraint ˆθ ML = arg max θ Θ 1 n n ln P (a i x i ; θ) i=1 s.t. P = Ψ(θ, P) Game-theoretic models are also characterized by a fixed point. Katsumi Shimotsu, Ken Yamada DP models JEA 2011 tutorial session 19 / 42

20 Game-theoretic models Many markets are characterized as a competition among a few firms with differentiated products. In such markets, strategic interaction between firms become important. We want to understand how firms compete with each other in such markets. Implications for competition policy, entry regulation, etc. Structural economic models" explicitly model firms strategic interaction (and dynamic choice) Katsumi Shimotsu, Ken Yamada DP models JEA 2011 tutorial session 20 / 42

21 Simple static game of entry Two potential entrants to a market (for example, 5th generation smartphone, Docomo and KDDI). If both firms enter, both firms earn positive profit. If only one firm enters, it earns larger profit. Payoff matrix firm 2 in out firm 1 in (1, 1) (3, 0) out (0, 3) (0, 0) Nash equilibrium: given the other player s action, I have no incentive to change my action (in,in) Katsumi Shimotsu, Ken Yamada DP models JEA 2011 tutorial session 21 / 42

22 Games with private information But firm 1 may not know everything about firm 2. Payoff matrix with a random component firm 2 in out firm 1 in (ε 1 θ, ε 2 θ) (ε 1, 0) out (0, ε 2 ) (0, 0) ε 1 and ε 2 are drawn independently from uniform distribution[0, 1]. Both firms know θ. Only firm 1 observes ε 1, and only firm 2 observes ε 2 (private information). Katsumi Shimotsu, Ken Yamada DP models JEA 2011 tutorial session 22 / 42

23 How to determine the equilibrium firm 2 in out firm 1 in (ε 1 θ, ε 2 θ) (ε 1, 0) out (0, ε 2 ) (0, 0) Suppose firm 1 thinks that firm 2 enters with probability P 2 = firm 1 s subjective probability ( belief") Firm 1 enters the market if P 2 (ε 1 θ) + (1 P 2 )ε 1 > 0 ε 1 > θp 2 Because ε 1 U[0, 1], firm 1 enters the market with probability 1 θp 2 when his belief is P 2. Katsumi Shimotsu, Ken Yamada DP models JEA 2011 tutorial session 23 / 42

24 How to determine the equilibrium Firm 1 enters the market with probability 1 θp 2 when his belief is P 2. Firm 2 enters the market with probability 1 θp 1 when his belief is P 1. Bayesian perfect equilibrium: both players belief must be consistent with their action. P 1 = 1 θp 2, P 2 = 1 θp 1 P 1 = P 2 = 1/(1 + θ) Best response mapping: [0, 1] 2 [0, 1] 2 Ψ(θ, P) = (1 θp 2, 1 θp 1 ). Then BPE is a fixed point of Ψ : P = Ψ(θ, P ) Katsumi Shimotsu, Ken Yamada DP models JEA 2011 tutorial session 24 / 42

25 Estimation of θ in this model Suppose we have iid data of the entry decision (a 1i, a 2i ) for i = 1,..., n, and we want to estimate the value of θ. If we assume the data are in the equilibrium, we can estimate θ by [ ˆθ = 1 1 ˆP ˆP ] 2. 2 ˆP 1 ˆP 2 Katsumi Shimotsu, Ken Yamada DP models JEA 2011 tutorial session 25 / 42

26 Dynamic discrete game N firms = potential entrants Entry/exit choice: a it A = {0, 1}. Firm i s profit in period t: Π i (a t, S t, a i,t 1, ɛ t ; θ) All the firms current decision: a t = (a 1t,..., a Nt ) Market demand condition: St (observable) Past entry decision: a i,t 1 Private shocks: ɛt = (ɛ 1t,..., ɛ Nt ). θ: parameter of interest Katsumi Shimotsu, Ken Yamada DP models JEA 2011 tutorial session 26 / 42

27 Dynamic discrete game Profit function of firm i: Π i (a t, S t, a t 1, ɛ it ; θ) = θ RS ln S t θ FC,i θ EC (1 a i,t 1 ) θ RN ln(1 + j i a jt ) + ɛ it1 θ RS : Revenue parameter θ FC,i : Operating cost θ EC : Entry cost θ RN : Degree of strategic substitution Katsumi Shimotsu, Ken Yamada DP models JEA 2011 tutorial session 27 / 42

28 Dynamic discrete game Dynamic optimization by firm i [ ] E β t Π i (a t, S t, a i,t 1, ɛ t ; θ) S t, a t 1 ; θ max a i1,a i2,... t=0 Assume the state variable follows a Markov process Markov decision problem given his belief Stationary solution Katsumi Shimotsu, Ken Yamada DP models JEA 2011 tutorial session 28 / 42

29 Dynamic discrete game Empirical implication: firm i s conditional choice probabilities = P i (a x). P i = {P i (a x)} (a,x) : firm i s conditional choice probabilities for all possible x The conditional choice probabilities of all the firms: P = (P 1,..., P N ) For a given θ, an equilibrium is characterized by a fixed point of the best response mapping P = Ψ(θ, P) We assume Ψ(θ, P) has a unique fixed point for the moment. Katsumi Shimotsu, Ken Yamada DP models JEA 2011 tutorial session 29 / 42

30 P can be large For example, 5 potential entrants all the firms entry status in the previous period: 2 5 = 32 support points market condition takes 10 different values x t takes = 320 different values Length of P = = 1600 finding a fixed point of Ψ can be computationally costly Katsumi Shimotsu, Ken Yamada DP models JEA 2011 tutorial session 30 / 42

31 Economic models with a fixed point constraint When P = P(a x) is the choice probability of a discrete action a conditional on x, the log-likelihood function is Q n (θ) = n ln P(a i x i ) s.t. P = Ψ(θ, P) i=1 Katsumi Shimotsu, Ken Yamada DP models JEA 2011 tutorial session 31 / 42

32 Constraint optimization approach (Su and Judd, 2012) Write down the Lagrangian L(θ, P, λ) = 1 n n ln P (a i x i ) + λ(p Ψ(θ, P)) i=1 Solve the first-order condition θ L(θ, P, λ ) = 0, P L(θ, P, λ ) = 0, P Ψ(θ, P ) = 0 According to the authors, one can solve this problem using the NEOS Server, a free internet service which gives the user access to several state-of-the-art solvers." Katsumi Shimotsu, Ken Yamada DP models JEA 2011 tutorial session 32 / 42

33 NPL estimator (Aguirregabiria and Mira, 2007) P 0 : initial guess of P Step 1: Given P k 1, 1 n n ln[ψ(θ, P k 1 )](a i x i ). i=1 can be viewed as a pseudo log-likelihood function. So, estimate θ by maximizing this objective function θ k Step 2: Given θ k, update the estimate of P by P k = Ψ( θ k, P k 1 ). Iterate Steps 1-2: { θ k, P k } k=1 Katsumi Shimotsu, Ken Yamada DP models JEA 2011 tutorial session 33 / 42

34 NPL Fixed Points: {ˇθ, ˇP} If the sequence { θ k, P k } k=1 converges, its limit satisfies ˇθ = arg max θ ˇP = Ψ(ˇθ, ˇP). 1 n n ln[ψ(θ, ˇP)](a i x i ), The NPL estimator is defined as (ˇθ, ˇP) that achieves the highest pseudo-likelihood value. Easy to implement: standard optimization and policy iteration i=1 Katsumi Shimotsu, Ken Yamada DP models JEA 2011 tutorial session 34 / 42

35 Convergence of the NPL algorithm? Convergence of { θ k, P k } k=1? Empirical researchers report non-convergence of the NPL algorithm. For example, it can exhibit a 2-period cycle. Katsumi Shimotsu, Ken Yamada DP models JEA 2011 tutorial session 35 / 42

36 Property of NPL updating (Kasahara and Shimotsu, 2012) In a neighborhood of P 0, θ j ˆθ NPL = O p ( P j 1 ˆP NPL ), P j ˆP NPL = M Ψθ Ψ P ( P j 1 ˆP NPL ) + smaller order terms, where Ψ P = P Ψ(θ, P) = Jacobian of Ψ(θ, P) M Ψθ = I Ψ θ (Ψ θ P Ψ θ ) 1 Ψ θ P The convergence of P k depends on the dominant eigenvalue of M Ψθ Ψ P. Katsumi Shimotsu, Ken Yamada DP models JEA 2011 tutorial session 36 / 42

37 Dynamic Game Example (continued) Profit function of firm i: Π i (a t, S t, a t 1, ɛ it ; θ) = θ RS ln S t θ FC,i θ EC (1 a i,t 1 ) θ RN ln(1 + j i a jt ) + ɛ it1 θ RS : Revenue parameter θ FC,i : Operating cost θ EC : Entry cost θ RN : Degree of strategic substitution Katsumi Shimotsu, Ken Yamada DP models JEA 2011 tutorial session 37 / 42

38 Dominant eigenvalue of Ψ P and M Ψθ Ψ P θ RN ρ(ψ P ) ρ(m Ψθ Ψ P ) Note: dim(p) = 144 and dim(θ) = 2. Strong strategic substitutability { θ k, P k } diverges away from (θ 0, P 0 ). Katsumi Shimotsu, Ken Yamada DP models JEA 2011 tutorial session 38 / 42

39 What if Ψ is not locally contractive around P 0? (Kasahara and Shimotsu, 2012) (1) Relaxation method: [Λ(θ, P)](a x) = {[Ψ(θ, P)](a x)} α P(a x) (1 α), α (0, 1) easy, works in some cases (2) Recursive Projection Method Katsumi Shimotsu, Ken Yamada DP models JEA 2011 tutorial session 39 / 42

40 References I Aguirregabiria, V. and P. Mira (2002). Swapping the nested fixed point algorithm: a class of estimators for discrete Markov decision models. Econometrica 70, Aguirregabiria, V. and Mira, P. (2007). Sequential estimation of dynamic discrete games. Econometrica 75, Aguirregabiria, V. and Mira. P. (2010). Dynamic discrete choice structural models: A survey. Journal of Econometrics 156, Hotz, J. and R. A. Miller (1993). Conditional choice probabilities and the estimation of dynamic models. Review of Economic Studies 60, Imai, S., Keane, M. P. (2004). Intertemporal labor supply and human capital accumulation. International Economic Review 45, Katsumi Shimotsu, Ken Yamada DP models JEA 2011 tutorial session 40 / 42

41 References II Kasahara, H. and K. Shimotsu (2008) Pseudo-likelihood Estimation and Bootstrap Inference for Structural Discrete Markov Decision Models. Journal of Econometrics 146, Kasahara, H. and K. Shimotsu (2012) Sequential Estimation of Structural Models with a Fixed Point Constraint. Econometrica 80, Keane, M. P. and K. I. Wolpin (1997). The Career Decisions of Young Men. Journal of Political Economy, 105, Rust, J. (1987). Optimal replacement of GMC bus engines: an empirical model of Harold Zurcher. Econometrica 55, Rust, J., Rothwell, G. (1995). Optimal Response to a Shift in Regulatory Regime: The Case of the US Nuclear Power Industry. Journal of Applied Econometrics 10, S75-S118. Katsumi Shimotsu, Ken Yamada DP models JEA 2011 tutorial session 41 / 42

42 References III Rust, J., Phelan. C. (1997). How Social Security and Medicare Affect Retirement Behavior In a World of Incomplete Markets. Econometrica 65, Su, Che-Lin and Judd, K. L. (2012). Constrained optimization approaches to estimation of structural models. Econometrica 80, Katsumi Shimotsu, Ken Yamada DP models JEA 2011 tutorial session 42 / 42